Clinical Ultrasound Examination Dataset

• Clinical Ultrasound Examination Datasets are curated collections of real B-mode images acquired under standardized protocols to support diagnostic and algorithmic research.
• They include multi-plane scans, associated pose calibration files, and rich acquisition metadata, enabling quantitative evaluations using metrics like PSNR, SSIM, and MSE.
• The open-access nature and structured design facilitate reproducible benchmarking, pretraining for rendering models, and further clinical research applications.

A clinical ultrasound examination dataset is a systematically curated collection of ultrasound image data, typically acquired under defined clinical protocols and annotated to support a range of diagnostic, quality control, or machine learning benchmarks. These datasets may consist of static B-mode frames or cine video sequences, cover diverse anatomical regions, and encode pertinent clinical or technical metadata, making them essential resources for algorithmic development, quantitative analysis, and evaluation of automated or computer-assisted ultrasound assessment methodologies.

1. Dataset Scope and Clinical Protocols

Clinical ultrasound examination datasets originate from routine or research-driven imaging workflows in hospitals and specialty labs. Acquisitions are performed according to standardized sonographer protocols tailored to the target anatomy—for example, deep or superficial linear scanning for extremities (wrist, carpal tunnel), curvilinear probe sweeps for abdominal organs (kidney, liver), or specialized planes for cardiac echocardiography. Institutional protocols dictate probe choice (e.g., high-frequency linear for wrist scans, convex for kidney), imaging parameters (e.g., gain, depth, focal zones), and anatomical coverage (transverse or longitudinal sweeps, inclusion of full cross-section). The Clinical Ultrasound Examination Dataset described in UltraGS utilizes a Canon i900 system and covers wrist and kidney exams, each with multi-plane sweeps under real-world clinical guidance (Yang et al., 11 Nov 2025).

2. Data Composition and Organization

A typical clinical examination set comprises multiple cases (patients or paper subjects), each associated with a series of B-mode ultrasound frames. In the UltraGS Clinical Dataset, six cases (three wrist, three kidney) are provided, with each sweep likely containing upwards of 80 frames or more (exact counts are to be reported upon release) (Yang et al., 11 Nov 2025). Datasets structure images in case/organ-specific directories, with standardized file naming and paired metadata. For example:

/ClinicalDataset/ Case1_wrist/ images/ img_00001.png, ..., img_N.png poses/ cam_00001.txt, ..., cam_N.txt ... Case6_kidney/ ...

Image resolution typically mirrors clinical scanner output (e.g., 512×512 or 600×800 pixels, 8-bit PNG), while pose and intrinsic/extrinsic calibration files (e.g., COLMAP FOV and [R|t] matrices) enable post-hoc spatial modeling or 2D/3D reconstruction required for downstream analysis and view synthesis.

3. Metadata, Annotation, and Quality Dimensions

Clinical datasets systematically encode scan-level and frame-level metadata to enable comprehensive downstream analysis. This includes probe type, pixel spacing, depth and gain settings, exam date, operator identifier, and clinical notes (e.g., anatomical region scanned). In UltraGS, each frame is associated with pose metadata necessary for 2D Gaussian splatting and depth-aware rendering, supporting tasks such as novel view synthesis where the camera model is paramount (Yang et al., 11 Nov 2025). Unlike some datasets that include segmentation masks or diagnostic labels (e.g., lesion bounding boxes in breast cancer, tissue masks in kidney segmentation (Lin et al., 3 Jun 2024, Singla et al., 2022)), the Clinical Ultrasound Examination Dataset is “unlabeled B-mode”—no ground-truth anatomical masks or pathology annotations are supplied, focusing on acquisition realism for physics-based rendering and algorithmic benchmarking.

4. Benchmarking and Quantitative Evaluation

Clinical ultrasound datasets form the basis for objective benchmarking of machine learning models and image synthesis algorithms. Evaluation metrics depend on task: for image reconstruction/synthesis, conventional quantitative measures include:

• Mean Squared Error (MSE):

$$\text{MSE}(I, \hat{I}) = \frac{1}{WH} \sum_{x=1}^{W} \sum_{y=1}^{H} (I(x,y) - \hat{I}(x,y))^2$$

• Peak Signal-to-Noise Ratio (PSNR):

$$\text{PSNR} = 10 \log_{10} \left( \frac{\text{MAX}_I^2}{\text{MSE}} \right)$$

where $\text{MAX}_I$ is the maximum intensity value (e.g., 1.0 after normalization).

• Structural Similarity Index (SSIM):

$$\text{SSIM}(I, \hat{I}) = \frac{(2\mu_I\mu_{\hat{I}} + c_1)(2\sigma_{I\hat{I}} + c_2)}{(\mu_I^2 + \mu_{\hat{I}}^2 + c_1)(\sigma^2_I + \sigma^2_{\hat{I}} + c_2)}$$

UltraGS demonstrates its depth-aware Gaussian splatting method using these metrics on the Clinical Dataset, reporting state-of-the-art performance (e.g., PSNR up to 29.55, SSIM up to 0.89, and MSE as low as 0.002) (Yang et al., 11 Nov 2025). Frame selection for benchmarking follows a protocol (e.g., every 8th frame selected for test split), but no dedicated validation set is specified.

5. Data Access, Licensing, and Practical Utilization

Clinical ultrasound datasets are increasingly released under open-access licenses to facilitate reproducibility and global research collaboration. The UltraGS Clinical Dataset is made available via GitHub under an MIT-style license for academic and non-commercial research use; adherence to ethical guidelines and exclusion of patient identifiers is ensured (Yang et al., 11 Nov 2025). Downloadable assets encompass image frames, calibration/pose files, and case-level metadata. No further preprocessing or denoising is applied beyond intensity normalization to [0,1]; researchers may apply custom speckle reduction or filtering contingent on methodological requirements.

Standard usage includes:

• As a pretraining or benchmarking set for physics-based view synthesis and rendering frameworks.
• For development and evaluation of deep learning methods requiring pose-labeled real clinical B-mode imagery.
• As a template for extending data acquisition to additional organs or dynamic scanning protocols using the same data structure and metadata conventions.

6. Methodological Significance and Research Frontiers

The publication of clinical ultrasound examination datasets such as that in UltraGS (Yang et al., 11 Nov 2025) addresses the chronic bottleneck posed by absence of real-case, pose-labeled imagery in novel-view synthesis, spatial reconstruction, and clinical AI research. The provision of detailed pose metadata enables rigorous evaluation of geometric and photometric rendering strategies, establishing a new paradigm for benchmarking beyond purely diagnostic or segmentation-focused datasets. Limitations include the small subject number (six cases) and the absence of labeled ground truth or clinical findings, restricting certain supervised learning paradigms. However, the structure and licensing facilitate rapid extension with additional labeled or annotated content to meet broader clinical and algorithmic validation needs.

In summary, a clinical ultrasound examination dataset is defined by its origin in real clinical scanning protocols, organization by anatomical region and frame, comprehensive metadata pairing, open academic access, and suitability for quantitative benchmarking across a spectrum of computer vision, machine learning, and image analysis tasks (Yang et al., 11 Nov 2025).